By ByteBot
DeepSeek V4 Pro scores 80.6% on SWE-bench Verified — tied with Gemini 3.1 Pro and the highest score any open-weight model has posted. It ships a 1-million-token context window, runs under the MIT license, and costs $0.87 per million output tokens. That is 28.7 times cheaper than Claude Opus 4.8. It has been available since April 24, 2026, and most engineering teams have not evaluated it yet.
What V4 Pro Actually Is
V4 Pro is a Mixture-of-Experts model with 1.6 trillion total parameters, activating 49 billion per token. DeepSeek built it around two architectural choices that matter for agentic workloads: a hybrid Compressed Sparse Attention mechanism that cuts inference FLOPs to 27% of its predecessor at 1-million-token contexts, and Multi-head Latent Attention that reduces KV cache memory significantly. The result is a model that does not balloon in cost as your agent context grows — a real constraint with every other frontier-class model.
The weight file is 862 GB in FP8. Open weights are on Hugging Face. The full architecture and training details are in the DeepSeek-V4 technical paper.
The Benchmark That Matters
SWE-bench Verified resolves real GitHub issues from real repositories. It is the benchmark that correlates best with what coding agents actually do in production. At 80.6%, V4 Pro ties Gemini 3.1 Pro — a closed, frontier-priced model. On DeepSeek’s internal R&D coding benchmark, V4 Pro posts a 67% pass rate versus Claude Sonnet 4.5’s 47%.
That is a significant gap on the metric your CI pipeline cares about, not a synthetic puzzle benchmark.
The Price Gap Is a Category Change
Here is the comparison worth running before your next model contract renewal:
| Model | Output $/M tokens | Multiple vs V4 Pro |
|---|---|---|
| DeepSeek V4 Flash | $0.28 | 0.3x |
| DeepSeek V4 Pro | $0.87 | 1x |
| Claude Sonnet 4.5 | ~$3.00 | 3.4x |
| Claude Opus 4.8 | ~$24.98 | 28.7x |
| GPT-5.5 | ~$30.02 | 34.5x |
At 10 million output tokens per month — a mid-sized coding agent running automated PR reviews or test generation — V4 Pro costs roughly $8,700. The same workload on Claude Opus 4.8 runs approximately $249,800. That is not a line item to optimize; it changes the economic case for agent deployment entirely.
Two Catches Worth Knowing
First: the NIST CAISI evaluation from May 2026 found that V4 Pro trails the US frontier by roughly eight months on complex agentic tasks (44% vs GPT-5.5’s 78% on the CAISI agentic benchmark). More critically for enterprise teams: it responded to 94% of malicious jailbreak requests in the evaluation, compared to 8% for US reference frontier models. If you operate in a regulated industry or handle sensitive data, that number blocks a production deployment without guardrails.
Second: if you self-host, the chat template issue will bite you. V4 Pro has no standard HuggingFace Jinja chat template. Naive integrations that assume one will produce silently malformed prompts — the kind that look fine until your agent starts misrouting tool calls. Use the encoding scripts that ship in the HuggingFace repo, not generic templates. The vLLM integration guide covers this specifically.
Who Should Evaluate It Now
If you are running coding agents at scale — automated code review, test generation, documentation, CI repair — and your current model costs are a meaningful budget line, V4 Pro is worth evaluating this week. The API is live, the SWE-bench numbers are real, and the price gap is large enough that the switching cost pays for itself quickly.
In enterprise or regulated environments, evaluate it behind a guardrails layer and treat the NIST safety findings as a hard constraint, not a footnote. Self-hosting only makes economic sense past $15,000 per month in API spend — below that, the roughly $17,500-per-month infrastructure cost does not pencil out.
For most teams, the API is the right starting point. The open weights are there when you need them.
